Why pre-AI business archives are becoming more valuable to AI developers

Why do pre-AI business records carry a premium?

Records a company created before generative AI tools went mainstream at the end of 2022 can be shown to be human-made, and that matters to the people building AI. As more new email, code and documents are drafted with AI assistance, developers increasingly want material that clearly reflects human judgment, with the decisions and consequences that followed.

Picture an engineering firm's 2014 log of requests for information. Each entry holds a contractor's question, the engineer's reply, the drawing revision that resulted and the change order that came later. No model wrote any of it. A company with ten or more years of records like that holds something a newer business cannot manufacture.

What is the low-background steel analogy?

Steel made before the first nuclear weapon tests carries less radioactive contamination than steel made since, so it is prized for instruments that must measure very low levels of radiation. People who work on AI data borrow the idea for text produced before AI writing tools spread: material free of machine-generated content.

The analogy has limits. Not every pre-AI record is valuable, and recent records are far from worthless. But it explains why provenance now carries weight. Developers are wary of training on large volumes of machine-generated text, which is one reason data with a clearly human origin is attractive. This page cites no study on that point.

How scarce is human-made text?

Public human writing is finite. Researchers at Epoch AI estimated the effective stock of human-generated public text at roughly 300 trillion tokens and projected that, if current trends continue, language models will fully use it between 2026 and 2032 (Epoch AI analysis). It is a forecast with wide uncertainty, but it explains why attention has turned to non-public sources, company archives among them.

Company records add something the public web lacks: they show work being done, not just writing about work. The wider gap is described in the work that rarely reaches AI training sets, and the limits of manufactured substitutes in how synthetic environments compare with real business logs.

Which archives count as pre-AI data?

Almost any system that has run for years can hold pre-AI records. The question is whether the history survived.

Email is often the deepest of these; why business email archives are valuable for AI explains what makes it useful.

The archive provenance check

Work through this list with whoever runs your IT, whether that is an internal team or an outside provider.

• We know how far back each main system goes, and someone can confirm it.
• Past migrations carried history across, or the old system's export was kept.
• Retired systems, former employees' mailboxes and backups still exist and can be restored.
• Original timestamps and metadata survive, not just re-saved copies.
• Our own people created these records in our own work; we are not holding them for clients or bought them from others.
• Nothing in the archive was generated with AI to pad it out.
• We are a US company with 50+ full-time employees at peak (contractors excluded), several years of documented operations, and an owner or executive who can authorize a license.

If most boxes are ticked, answer the short questions in the company fit checker for an early, non-binding indication; it does not ask who you are.

What raises or lowers the value of an old archive?

Age helps only when the records are intact, connected and yours.

Which rights questions do older archives raise?

Old records were collected under old promises, and those promises still count.

• Privacy commitments. FTC staff warned in February 2024 that it may be unfair or deceptive for a company to adopt more permissive data practices, such as using data for AI training, and tell people only through a quiet, retroactive change to its terms or privacy policy (FTC staff post). Check what customers were told when the records were collected.
• Client contracts. Confidentiality clauses signed years ago may limit what can be licensed today.
• Personal content. Old mailboxes and chat channels contain personal messages, so the company sets masking and redaction rules for that content before anyone touches the archive.
• Acquired businesses. Archives that came with an acquisition depend on what the purchase documents transferred.

This is general information, not legal, tax or financial advice. Confirm with your own counsel before acting.

Why does timing matter for owners?

Archives tend to disappear at transitions: a sale, a system consolidation, a wind-down or a retirement. McKinsey reports that more than half of US small-business owners are over 55 and that about six million US small and medium-size businesses will face ownership transitions by 2035 as baby boomers retire (McKinsey, The great ownership transfer). For an owner thinking about the next five years, the practical rule is simple: before any system is retired, keep a complete export.

Companies that are still operating, have been acquired or have wound down can all qualify if the data still exists. Some owners worry about what models trained on their records might do; whether AI trained on a company's records could replace it is answered directly elsewhere.

How does an owner take this forward?

The process is built so the company stays in control at every step.

1. Run the fit checker, then apply directly at sourcex.si/apply, or ask the adviser who raised the idea to introduce you with their referral link.
2. SourceX confirms headcount, years of history, systems and rights with you.
3. Your team lists each system and the years it covers in a data inventory.
4. You agree one all-in price and the terms, with SourceX's fee included and no separate charges; nothing is binding until you sign.
5. Buyers, meaning AI labs and other data buyers, look at the opportunity and typically respond within about two weeks once the company is deal-ready.
6. After signing, the agreed records are prepared under the agreed redaction rules and delivered, and you receive a one-time payment, which typically arrives within about 60 days of invoicing after the buyer has chosen the data.

You keep ownership throughout; the data is licensed, not sold. More detail is in how it works.

Owners and advisers who know other established companies can also refer them. Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company, and the reward becomes payable only after the buyer pays and SourceX receives its fee.

Next step

If your archives reach back well before 2022, apply at sourcex.si/apply. If you advise owners with long histories, register as a partner and make the introduction.